ROAD

ROAD: Reciprocal-Objective Alignment of Discriminative Semantics for 3D Shape Generation

ROAD is a 3D shape generation model that transfers discriminative 3D semantics into the generative process through global feature alignment and token-level matching. It is built on the Step1X-3D rectified-flow framework and uses a frozen Uni3D teacher during alignment training.

Task

ROAD is designed for conditional 3D shape generation from visual inputs. Because this task does not currently map cleanly to a dedicated Hugging Face pipeline, the repository uses the supported other pipeline tag.

Input and Output

Input

• One or more reference images representing an object
• Optional model and sampling configuration

Output

• A generated 3D object representation
• The resulting geometry can be exported or processed as a 3D mesh, such as a GLB asset

Repository Contents

This repository contains model files and related resources for ROAD. The accompanying source repository provides:

• ROAD and baseline training configurations
• Step1X-3D-based training code
• Uni3D global feature alignment
• Token-level Hungarian matching
• Uni3D-I and ULIP-I evaluation tools
• Data preparation and training scripts

Model files in this repository should be used with the implementation and setup instructions provided in the GitHub project.

Usage

Download this repository with the Hugging Face CLI:

huggingface-cli download H-EmbodVis/ROAD --local-dir pretrained/road

Alternatively, clone it with Git LFS:

git lfs install
git clone https://huggingface.co/H-EmbodVis/ROAD

Place the downloaded files in the location expected by the ROAD configuration, then follow the training or inference instructions in the source repository.

Source Code

• GitHub: https://github.com/H-EmbodVis/ROAD
• Project page: https://h-embodvis.github.io/ROAD/
• Paper: https://arxiv.org/abs/2607.28581

License

The Step1X-3D-derived ROAD training code is distributed under the Apache License 2.0. Components derived from Uni3D and ULIP retain their respective upstream licenses. External checkpoints and datasets may be subject to separate license terms.

Citation

If you use ROAD in your research, please cite:

@misc{luo2026roadreciprocalobjectivealignmentdiscriminative,
  title         = {ROAD: Reciprocal-Objective Alignment of Discriminative Semantics for 3D Shape Generation},
  author        = {Xiao Luo and Mingyang Du and Xin Zhou and Tianrui Feng and Xiwu Chen and Xiaofan Li and Jiangning Zhang and Dingkang Liang},
  year          = {2026},
  eprint        = {2607.28581},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.28581}
}

Acknowledgements

ROAD builds upon Step1X-3D, Uni3D, and ULIP. Please also follow the citation and licensing requirements of these upstream projects.